AI Disclosure
Synthia is built on AI. We believe in full transparency about what AI can and cannot do, and exactly how our product uses it.
Important: All Synthia research outputs are AI-generated synthetic data. They are intended for directional research and cultural sensitivity checks — not as a replacement for real user research. Always validate significant findings with real users before major product decisions.
What is AI-generated in Synthia?
Everything in a Synthia research report is AI-generated. This includes: user personas (names, ages, cultural backgrounds, digital behaviours), task responses (how each persona interacts with your design), friction points (usability issues identified by AI simulation), cultural flags (cultural mismatches between your design and the target market), and recommendations (prioritised actions to improve cultural fit).
No real users were recruited, interviewed, or observed to produce these outputs. All personas are fictional constructs built from cultural and demographic research data.
How are personas generated?
Personas are generated using large language models (LLMs) conditioned on demographic parameters (age range, gender distribution, location, language), cultural parameters (religiosity, family structure, decision-making style, digital literacy tier), and market-specific training context (e.g. UK South Asian diaspora, Gulf Arab, Nigerian urban).
The LLM simulates how a person with these characteristics would interact with a described UI, what friction they would encounter, and how they would feel about the product.
What AI models does Synthia use?
Synthia uses Together AI's hosted inference API, running Llama-3.3-70B as the primary model for persona simulation and report generation. Cultural calibration prompts are proprietary to Synthia and are not disclosed.
What are the limitations?
AI-generated research has significant limitations:
1. No real users. Synthia cannot capture the full complexity of real human behaviour, emotion, or context. AI personas do not have lived experience.
2. Western bias in base models. LLMs are predominantly trained on English-language, Western internet content. Synthia works to counteract this with cultural calibration, but bias cannot be fully eliminated.
3. Statistical non-representativeness. A panel of 10 AI personas is not equivalent to a statistically valid sample of real users. Do not use Synthia outputs in place of properly sampled quantitative research.
4. No real usability data. Synthia simulates behaviour — it does not capture actual click paths, eye tracking, or real task completion times.
Synthia is appropriate for early-stage directional research, cultural sensitivity checks, and rapid market exploration. It is not appropriate as the sole basis for major product decisions.
How we label AI-generated content
All Synthia reports include a disclaimer that outputs are AI-generated synthetic research. We encourage users to share this context with stakeholders when presenting findings.
Data and model training
Your study inputs (market selection, research questions, design descriptions) are sent to Together AI for inference but are not used to train Synthia's models or Together AI's models. See Together AI's privacy policy for their data handling practices.